Real CLI stalled device-token poll proof Run driver.mjs with checkout path, side (before/after), and an unused local port. Source the private T3 Connect test-stage environment first. Requires Node 26, installed checkout dependencies, and preload.mjs beside driver.mjs. Example: node driver.mjs before 18431 The driver invokes the checkout's real apps/server/src/bin.ts command: connect login --headless --base-dir The preload routes only OAuth device authorization/token fetches through a local HTTP fault proxy. The proxy forwards unmodified requests and responses to the real Clerk test stage. It does not fabricate authorization responses or tokens. The first device-token request >=305 seconds after device-code issuance is held without response headers. All earlier/later requests forward normally. Open each private authorization.json verification_uri_complete in the browser, sign into the supplied stage account, and leave Approve unclicked until HOLD. Click Approve while that request is held. Both runs use this same interaction. Browser approval recordings supplement the timestamped CLI/proxy recordings. record.py records a live 1280x800 terminal-style viewer of sanitized actual CLI stdout and proxy events at 1 fps; it is not a recording of a native terminal. Recording began during the pending-poll phase, before either held poll. The events transcripts retain the complete run, including startup. No time acceleration or fake clock is used. No repository code is changed. Expected on base: no retry before device-code expiry, exit 1 despite approval. Expected on head: held connection closes after 10 seconds, next request follows 10 seconds later (5-second polling interval increased by 5 seconds), exit 0. Private CLI output, device authorization JSON, home state, credentials and tokens must not be uploaded. Uploaded transcripts redact verification links/user codes, account identity and checkout/home paths.